Not so long ago, corporations encouraged employees to use artificial intelligence without restrictions — tokens were handed out like corporate bonuses, and internal rankings held up those who "burned" the most as examples. Today, the picture is radically different: managers are hastily revoking access to models and imposing strict limits, revising budgets. How did tokens go from free fuel to the main line item of unforeseen expenses?

A token in the context of AI is the minimum unit of data processed by a language model. One token can correspond to part of a word, a whole word, or even a character. On average, 100 words break down into 130–140 tokens. They are divided into input (user query) and output (model response). For a long time, this unit was merely a technical abstraction until March 2023, when OpenAI released GPT-4, setting public rates: $30 per 1 million input tokens and $60 per 1 million output tokens. Monetization began from that moment.

From 2023 to 2025, the "era of subsidization" lasted: regular users paid via subscriptions, while corporate developers paid for tokens at soft prices. However, since spring 2026, all major providers have switched to full monetization of queries. First Anthropic, then GitHub Copilot. We are now witnessing the beginning of the "tokenapocalypse" — a period when unforeseen costs for the tokens themselves are added to spending on computing power.

The Price and Demand Paradox

The paradox is that over three years, the average price of an input token has fallen by about 98%. Yet enterprise spending on AI has skyrocketed by 320% — from $1.2 million in 2024 to $7 million in 2026. The reason is the explosive growth in usage volumes, especially due to agentic technologies. While a simple linear query in 2023 cost about $0.04, a query to a full-fledged agent system already costs $1.2 — 30 times more. For complex workflows, the difference can reach 500 times. Economists call this the Jevons paradox: a decrease in the price of a resource leads to an increase in its consumption, but in the case of AI, the effect has been hypertrophied.

Case Studies: From Uber to Microsoft

Uber, which actively implemented Claude Code and Cursor, found that its annual programming budget was exhausted by April. The CTO personally spent $1,200 on tokens during a single two-hour demonstration. The company introduced a limit of $1,500 per month per employee, whereas previously it encouraged using AI "as much as possible."

Microsoft granted thousands of developers access to Anthropic's Claude Code, but after six months, it scrapped the experiment, revoking most direct licenses. Internal surveys showed that engineers preferred the competitor's tool, but its excessive use likely impacted costs. Abandoning Claude is a compelling signal that the economics of corporate AI coding, at current token prices, do not work. The tools are so good that they are used constantly, and this becomes a financial problem.

Amazon is restructuring its partnership with Anthropic, switching from paying for compute hours to paying for tokens. This could increase costs, although the company denies it. For giants like Microsoft and Amazon, the "tokenapocalypse" is a manageable risk, but more modest firms are already facing financial difficulties.

Costly Mistakes and the Future

There are known cases where companies forgot to set limits and received bills for $500 million. One developer spent over $40,000 on tokens in a single month. Costs per developer increased 18.6 times over nine months, and most companies still cannot measure the return on these expenses.

Against this backdrop, the Linux Foundation presented the Tokenomics Foundation project — a new standard for "discipline" in the context of AI costs, similar to how FinOps regulates cloud service spending.

My comment: Tokens have become not just a unit of measurement, but a full-fledged line item in the corporate budget. The "tokenapocalypse" is not a crisis, but an inevitable stage of industry maturation. Companies that fail to learn how to manage this resource risk finding themselves in a financial trap. Those who implement a systematic approach to accounting and limiting will gain a competitive advantage. In the coming years, we will see tokens become as predictable an expense as cloud service subscriptions.